The Anthropic data leak has revealed how the company is preparing its next major model release while also dealing with rising security pressure. Anthropic said it is developing a new general-purpose system that it views as its most capable model so far, with testing already underway among early-access customers. Exposed draft materials identified the model as Claude Mythos and described it as completed in training, expensive to run and limited to a cautious rollout rather than a broad public launch.
The same leaked materials outlined a higher model tier called Capybara, positioned above Opus on intelligence, size and cost. Draft benchmark claims said it outperformed Claude Opus 4.6 across coding, academic reasoning and cybersecurity evaluations. Anthropic said the documents became publicly accessible because of human error tied to an external content management tool and that access to the data store was restricted after the issue was reported.
What gives the Anthropic data leak broader significance is the context around it. Anthropic has also alleged large-scale distillation attempts involving fraudulent Claude accounts and separate researchers disclosed a chained Claude attack called Claudy Day that combined prompt injection, data exfiltration and redirect flaws. Together, those issues show that frontier model releases are increasingly shaped by security readiness as much as raw performance.
